Scheduled dedicated resource pool (pay-as-you-go)

Updated at:

A scheduled dedicated resource pool allows you to plan GPU computing resources of specified specifications in advance and define the time windows in which the resources take effect. Within the time windows, dedicated instances with the specified quantity and GPU card type are allocated to functions. When the GPU inventory is sufficient, a scheduled dedicated resource pool provides GPU computing power during business peaks and automatically releases resources during inactive hours to implement pay-as-you-go billing and reduce costs.

Important

A scheduled dedicated resource pool does not lock GPU resources. Within the time windows specified by a scheduled policy, the system attempts to activate GPU instances, but whether activation succeeds depends on whether the GPU inventory is sufficient at that time. If the inventory is insufficient, instances cannot be launched, and functions still cannot obtain GPU computing power within the time windows.

Overview

Scheduled dedicated resource pools apply only to GPU functions and use the pay-as-you-go billing method. Unlike a subscription dedicated resource pool, which is resident 24/7 and locks resources, a scheduled dedicated resource pool allows you to specify the time windows in which the resources take effect (hereinafter referred to as a scheduled policy). The system attempts to activate GPU instances only within the time windows specified by the policy. Outside the time windows, resources are automatically released and are not billed.

After you bind a scheduled dedicated resource pool to a function and allocate dedicated instances, the function uses the dedicated instances to process requests within the time windows if the GPU inventory is sufficient, which significantly reduces cold starts. If the inventory is insufficient within a time window, instances cannot be launched and the function cannot obtain GPU computing power. Outside the time windows, the scheduled dedicated instances are released.

Note that within a time window, the maximum number of requests that a function bound with a scheduled dedicated resource pool can process concurrently is the number of allocated dedicated instances × the instance concurrency. Requests that exceed this limit are throttled. Requests within this limit receive real-time responses.

Use cases

A scheduled dedicated resource pool is recommended if your workload has the following characteristics:

  • Tidal workloads: The workload has obvious peaks and valleys. For example, high-concurrency requests occur during a fixed time slot every day, such as 09:00-18:00, with almost no traffic during other hours.

  • Cost-sensitive resource utilization: You do not want to pay for resources that are resident 24/7. You want to request GPU resources on demand only when your workload needs them and pay only for the actual usage duration.

  • GPU computing demand with acceptable uncertainty: GPU computing power and low cold starts are required during active hours, but instance launch failures caused by insufficient inventory are acceptable, and elastic instances or other fallback solutions are available.

  • Flexible and controllable costs: You want to reduce overall resource costs through pay-as-you-go billing while meeting the demand for GPU computing power during active hours.

Billing

Scheduled dedicated resource pools use pay-as-you-go billing. Fees are calculated by multiplying the resource specifications by the actual active duration, and are accurate to the second. Fees are incurred only within the time windows specified by the scheduled policy, and no fees are incurred outside the time windows.

Billable items

Fees of a scheduled dedicated resource pool are calculated based on resource specifications and usage duration. The following billable items are supported:

  • GPU memory fees: billed based on GPU memory (GB) × active duration (seconds).

  • vCPU fees: billed based on the number of vCPUs × active duration (seconds).

  • Memory fees: billed based on memory (GB) × active duration (seconds).

  • Disk fees: billed based on disk capacity (GB) × active duration (seconds).

In a scheduled dedicated resource pool, the purchasable disk quota per card ranges from 10 GB to 200 GB, in 10 GB increments. Disk pricing varies by region:

  • Regions: China (Hangzhou), China (Shanghai), China (Beijing), China (Ulanqab), and China (Shenzhen)

    Disk price: CNY 0.0000007/GB/second

  • Regions: US (Virginia), Germany (Frankfurt), and Singapore

    Disk price: CNY 0.0000008/GB/second

Unit prices

GPU memory unit price

Billable item

GPU card type

Region

Price (CNY/GB/second)

GPU memory

Tesla series (tesla.1)

China (Hangzhou), China (Shanghai), China (Beijing), and China (Shenzhen)

0.0001385

Ampere series (ampere.1)

China (Hangzhou), China (Shanghai), China (Beijing), and China (Shenzhen)

0.000106

Ada series (ada.1)

China (Hangzhou), China (Shanghai), China (Beijing), and China (Shenzhen)

0.0000714

Ada.2 series (ada.2)

China (Hangzhou), China (Shanghai), China (Beijing), and China (Shenzhen)

0.0001344

Ada.3 series (ada.3)

China (Hangzhou), China (Shanghai), China (Beijing), and China (Shenzhen)

0.000096

Blackwell series (blackwell.1)

China (Hangzhou), China (Shanghai), China (Beijing), and China (Shenzhen)

0.0001399

Hopper series (hopper.1)

China (Hangzhou), China (Shanghai), China (Beijing), and China (Shenzhen)

0.0001145

Hopper.2 series (hopper.2)

China (Hangzhou), China (Shanghai), China (Beijing), and China (Shenzhen)

0.0001086

Xpu.1 series (xpu.1)

China (Hangzhou), China (Shanghai), China (Beijing), and China (Shenzhen)

0.0000697

Ada series (ada.1)

China (Ulanqab)

0.0000714

Ada.2 series (ada.2)

China (Ulanqab)

0.0001344

Ada.3 series (ada.3)

China (Ulanqab)

0.000096

Blackwell series (blackwell.1)

China (Ulanqab)

0.000121

Hopper series (hopper.1)

China (Ulanqab)

0.0001145

Hopper.2 series (hopper.2)

China (Ulanqab)

0.0001005

Xpu.1 series (xpu.1)

China (Ulanqab)

0.00007

Ada series (ada.1)

Germany (Frankfurt)

0.0001106

Hopper series (hopper.1)

Germany (Frankfurt)

0.000159

Hopper.2 series (hopper.2)

Germany (Frankfurt)

0.0001634

Ada series (ada.1)

US (Virginia)

0.0000666

Hopper series (hopper.1)

US (Virginia)

0.000115

Ada series (ada.1)

Singapore

0.0000867

Hopper series (hopper.1)

Singapore

0.0001374

Hopper.2 series (hopper.2)

Singapore

0.000157

vCPU unit price

Billable item

GPU card type

Region

Price (CNY/vCPU/second)

vCPU

Tesla series (tesla.1), Ampere series (ampere.1), Ada series (ada.1), Ada.2 series (ada.2), Ada.3 series (ada.3), Blackwell series (blackwell.1), Hopper series (hopper.1), Hopper.2 series (hopper.2), and Xpu.1 series (xpu.1)

China (Hangzhou), China (Shanghai), China (Beijing), China (Shenzhen), and China (Ulanqab)

0.0000167

Tesla series (tesla.1), Ampere series (ampere.1), Ada series (ada.1), Ada.2 series (ada.2), Ada.3 series (ada.3), Blackwell series (blackwell.1), Hopper series (hopper.1), Hopper.2 series (hopper.2), and Xpu.1 series (xpu.1)

Germany (Frankfurt), US (Virginia), and Singapore

0.0000246

Memory unit price

Billable item

GPU card type

Region

Price (CNY/GB/second)

Memory

Tesla series (tesla.1), Ampere series (ampere.1), Ada series (ada.1), Ada.2 series (ada.2), Ada.3 series (ada.3), Blackwell series (blackwell.1), Hopper series (hopper.1), Hopper.2 series (hopper.2), and Xpu.1 series (xpu.1)

China (Hangzhou), China (Shanghai), China (Beijing), China (Shenzhen), and China (Ulanqab)

0.0000083

Tesla series (tesla.1), Ampere series (ampere.1), Ada series (ada.1), Ada.2 series (ada.2), Ada.3 series (ada.3), Blackwell series (blackwell.1), Hopper series (hopper.1), Hopper.2 series (hopper.2), and Xpu.1 series (xpu.1)

Germany (Frankfurt), US (Virginia), and Singapore

0.0000123

Disk unit price

Billable item

GPU card type

Region

Price (CNY/GB/second)

Disk

Tesla series (tesla.1), Ampere series (ampere.1), Ada series (ada.1), Ada.2 series (ada.2), Ada.3 series (ada.3), Blackwell series (blackwell.1), Hopper series (hopper.1), Hopper.2 series (hopper.2), and Xpu.1 series (xpu.1)

China (Hangzhou), China (Shanghai), China (Beijing), China (Shenzhen), and China (Ulanqab)

0.0000007

Tesla series (tesla.1), Ampere series (ampere.1), Ada series (ada.1), Ada.2 series (ada.2), Ada.3 series (ada.3), Blackwell series (blackwell.1), Hopper series (hopper.1), Hopper.2 series (hopper.2), and Xpu.1 series (xpu.1)

Germany (Frankfurt), US (Virginia), and Singapore

0.0000008

Manage scheduled dedicated resource pools

Create a scheduled dedicated resource pool

  1. Log on to the Function Compute console. In the left-side navigation pane, choose Elastic Management > Dedicated Resource Pool.

    • Method 1: Purchase separately. Choose Purchase Dedicated Resource Pool > Apply for Scheduled Dedicated Resource Pool (Pay-As-You-Go).

    • Method 2: Purchase from a subscription instance and associate the pool with the subscription instance. Go to the details page of a dedicated resource pool (subscription) instance. Choose Scheduled Scaling > Apply.

  2. On the scheduled scaling application page, configure the following parameters:

    Parameter

    Description

    GPU card type

    The GPU card series. Options include Tesla, Ampere, Ada, Blackwell, and Hopper.

    vCPU specification

    The vCPU specification.

    Memory specification

    The memory specification.

    Disk specification per card

    The disk capacity allocated per GPU card. The value ranges from 10 GB to 200 GB, in 10 GB increments.

    Number of cards

    The number of GPU cards.

    Time zone

    The reference time zone for executing the scheduled policy. Default value: Beijing time (UTC+8).

    Effective date range

    The date range in which the resources take effect, accurate to the minute. For example, if you set this parameter to 2026-08-04 09:10 - 2026-09-04 09:10, the daily time windows take effect only within this date range.

    Daily time windows

    The time windows in which the resources take effect each day, accurate to the hour. For example, if you set this parameter to 09:00-18:00, the resources remain active every day during this time slot. Multiple non-contiguous time slots can be set for a single day.

  3. After the configuration is complete, click Submit Application. The system triggers an approval process based on the configured policy. After the application is approved, the system attempts to activate GPU instances at the beginning of each daily time window within the effective date range. Whether activation succeeds depends on whether the GPU inventory is sufficient at that time. Activated instances are automatically released at the end of each time window.

Modify the scheduled policy

After a scheduled dedicated resource pool is created, you can adjust the scheduled policy at any time. In the list of scheduled dedicated resource pools, click Modify Scaling Policy on the right side of the target pool, modify the number of cards, time zone, effective date range, or daily time windows, and then click Submit Application. After the application is approved, the new time windows take effect in the next scheduling cycle.

Note
  • Different time windows of the same pool must not overlap.

  • Policy modifications do not affect instances that have been activated and are running. Activated instances are not released early before the current window ends.

Scale out

On the details page of the target scheduled dedicated resource pool, click Modify and adjust the number of GPU cards as prompted to scale out the pool.

Only the number of GPU cards can be increased. The GPU card type, vCPU, memory, and disk specifications cannot be changed.

Delete

If you no longer need a scheduled dedicated resource pool, you can delete it. On the details page of the scheduled dedicated resource pool, click Delete in the upper-right corner and confirm the deletion. The deletion is irreversible. Fees that have already been incurred are settled in the next billing cycle.